Resolving uncertainty on the fly: modeling adaptive driving behavior as active inference.

Area of Science:

• Computational Neuroscience
• Traffic Psychology
• Human-Computer Interaction

Background:

• Developing accurate simulated human driver models is crucial for autonomous vehicle (AV) development and evaluation.
• Existing models often lack computational rigor, are scenario-specific, or are black boxes lacking interpretability.
• There is a need for generalizable, interpretable computational models of adaptive human driving behavior.

Purpose of the Study:

• To propose a novel, interpretable computational model for adaptive human driving behavior.
• To explain how drivers manage uncertainty by integrating goal-seeking and information-seeking behaviors.
• To demonstrate the model's applicability in diverse driving scenarios requiring uncertainty management.

Main Methods:

• Developed a computational model based on active inference, a framework from computational neuroscience.
• Formulated adaptive driving as a policy selection process minimizing expected free energy.
• Applied the model to simulated driving scenarios involving occlusions and visual time-sharing.

Main Results:

• The active inference model successfully explains the trade-off between progress and caution in driving.
• The model demonstrates how uncertainty resolution is integrated into goal-seeking behavior.
• Human-like adaptive driving behaviors emerged in simulations of complex scenarios.

Conclusions:

• Active inference provides a principled and interpretable framework for modeling adaptive human driving.
• The proposed model offers a unified approach to understanding goal-directed and uncertainty-resolving driving behaviors.
• This work advances the development of more realistic and explainable human driver models for AV research.